Google DeepMind launches AlphaGenome 2: the AI model that finds disease genes in single cells
The new model reads DNA cell type by cell type and, according to DeepMind, finds nearly twice as many genetic variants as its predecessor. Free for research, not approved for clinical use.


Google DeepMind on Wednesday launched AlphaGenome 2, a new AI model that predicts how small changes in DNA play out inside specific cell types in the body. The previous model looked at averages across whole tissues. The new one zooms in on one cell type at a time and, according to DeepMind, finds nearly twice as many genetic variants that switch genes on and off.
The company published the announcement on its blog and on its official X account on Wednesday 11 November at 16:03 Norwegian time (15:03 UTC). The model is available from the same day for non-commercial research through the AlphaGenome API, and in private preview for commercial research.
From tissue to single cells
The heart, the brain and the immune system share the same DNA. Yet a single-letter change can cause disease in some cells and not in others. That is what AlphaGenome 2 is meant to help researchers understand.
The first AlphaGenome model arrived in 2025. It read a DNA sequence of up to one million letters and produced thousands of predictions for many tissues at once. DeepMind writes that this parallel approach spread the model's capacity thin and made it hard to capture differences between closely related cell types.
AlphaGenome 2 has been redesigned. The model takes both the DNA sequence and a specific biological condition, such as a cell type, as input and uses that condition throughout its architecture. It is trained on human and mouse data, including more than 107 million single-cell profiles from 68 tissues. Alongside gene expression and chromatin, it now also predicts DNA methylation, which DeepMind calls a widely requested readout.
The numbers DeepMind reports
On the TenK10K dataset, a large map of 28 immune cell types, AlphaGenome 2 identified 30 percent of causal variants, according to DeepMind, compared with 16 percent for the original model fine-tuned on single-cell data. For the direction of the gene effect, whether a gene is turned up or down, the figures were 43 versus 23 percent.
DeepMind also gives a concrete example from inflammatory bowel disease (IBD). A recent study found that a variant in a distant regulatory switch overactivates the ETS2 gene in immune cells in the gut. AlphaGenome 2 reached the same finding computationally and showed that the effect was largely confined to those cells.
The model does not need retraining for new experiments either. Researchers can upload data from just a few dozen cells and get predictions for a cell state the model has not seen before. In an example from a COVID-19 mRNA vaccine study, the model distinguished activated immune cells from resting ones and showed that a variant linked to severe COVID-19 disrupted gene activity more strongly in the activated cells.
All of these results are DeepMind's own. The announcement does not cite an independent review, and DeepMind does not state a price for the commercial preview.
Straight into drug discovery
Genomics PLC, a British company that uses genetics in drug discovery, says in the announcement that it will build AlphaGenome 2 into its Mystra platform. CEO Peter Donnelly calls the model "an impressive advance in the field". DeepMind has also precomputed predictions for known trait-associated variants across hundreds of cell and tissue conditions, so researchers do not have to run the heavy computations themselves.
DeepMind says AlphaGenome 2 will soon be available as skills in Google Antigravity, the company's agentic development platform. That is announced, not launched.
DeepMind stresses that neither AlphaGenome nor AlphaGenome 2 has been validated or approved for clinical or diagnostic use. The model is a research tool.
What leaders should know
- AI is moving into the core of biological research. For pharmaceutical companies, research hospitals and biotech groups, models like this could help decide which experiments are worth running. That is our assessment, not a promise from DeepMind.
- The licence governs use. Open access covers non-commercial research. Commercial use requires a place in a private preview, with no stated price or terms.
- Not for the clinic yet. The model does not give answers about individual patients. Anyone evaluating it in health research should treat its output as hypotheses to be tested in the lab.
In September DeepMind launched AlphaGenome Atlas, its map of predictions for 9 billion DNA variants. AlphaGenome 2 is the next step: from tissue to single cells.
Sources and media
- Google DeepMind on X, 11 November 2026
- Google DeepMind: AlphaGenome 2 – Deciphering how cells read DNA to better understand disease
Illustration: AI-generated with Grok Bot / hogby.ai